Proceedings of the Annual Meeting of the Cognitive Science Society

Recent large language models (LLMs) can achieve human-like performance on some reasoning tasks, yet it remains unclear whether similar outcomes reflect similar reasoning processes. We compared human think-aloud protocols and prompted GPT-4o chain-of-thought traces on five Zendo reasoning tasks involving relational and compositional rule discovery. Reasoning traces were segmented and coded into predefined cognitive-state categories to examine state usage, transition patterns, and reasoning flow. Although GPT-4o matched human performance on simpler tasks, humans achieved higher overall accuracy (–ï60% vs. –ï30%), with larger divergence on tasks requiring more complex relational structure. Process-level analyses showed that human reasoning involved broader state diversity and more frequent cross-state transitions, whereas GPT-4o exhibited more concentrated transition patterns and higher rates of repeated local loops. These findings suggest that process-level analyses can reveal systematic differences in observable reasoning dynamics beyond task accuracy, highlighting the value of cognitive-state and transition analyses for comparing human and model-generated reasoning.